DOI 10.1007/s11229-011-0030-6
Received: 30 September 2011 / Accepted: 30 September 2011 / Published online: 1 November 2011
Springer Science+Business Media B.V. 2011
Abstract The paper shows why and how an empirical study of fast-and-frugal heuristics can provide norms of good reasoning, and thus how (and how far) rationality
can be naturalized. We explain the heuristics that humans often rely on in solving
problems, for example, choosing investment strategies or apartments, placing bets in
sports, or making library searches. We then show that heuristics can lead to judgments
that are as accurate as or even more accurate than strategies that use more information
and computation, including optimization methods. A standard way to defend the use
of heuristics is by reference to accuracy-effort trade-offs. We take a different route,
emphasizing ecological rationality (the relationship between cognitive heuristics and
environment), and argue that in uncertain environments, more information and computation are not always better (the less-can-be-more doctrine). The resulting naturalism
about rationality is thus normative because it not only describes what heuristics people
use, but also in which specific environments one should rely on a heuristic in order
to make better inferences. While we desist from claiming that the scope of ecological
rationality is unlimited, we think it is of wide practical use.
Keywords
G. Gigerenzer (B)
Max Planck Institute for Human Development, Berlin, Germany
e-mail: gigerenzer@mpib-berlin.mpg.de
T. Sturm
Department of Philosophy, Universitat Autnoma de Barcelona, Bellaterra (Barcelona), Spain
e-mail: tsturm@mpiwg-berlin.mpg.de
123
244
123
245
theoretical and practical rationality, between means-ends (or instrumental) and principled rationality, between procedural and substantive rationality, between the rationality
of rules or the rationality of persons, and so on (Mele and Rawlings 2004; Hanna 2006),
although these are largely irrelevant here (but see Sect. 5).
The rationality naturalized by our psychological research program picks out one
specific sense of rationality, albeit an important one. We start from a kind of meansends rationalitythe default notion of rationality (Nozick 1993, p. 133): to know
by which means to achieve ones ends always requires empirical knowledge. But our
naturalism about rationality is not merely based upon this triviality. It takes a distinctive and substantive further step by claiming that (1) humans rely on heuristics to solve
problems, and (2) heuristics can lead to more accurate judgments than strategies that
use more information and computation, including optimization methods, if one takes
into account the relation between a reasoners heuristics and his or her environment.
We claim that normative theories of rationality have paid too little, if any, attention
to this important relation. Thus, we supplement the instrumental notion of rationality
by an ecological dimension. The crucial point to which this leads is that the claim
(2), which has strong empirical evidence in its favor, is not merely descriptively relevant. In an uncertain environment, more information and computation are not always
better, as we shall argue. This doctrine of less-is-moreor more precisely, lesscan-be-moreis the core of the current attempt to naturalize a particular domain of
normative theories of reasoning and decision making. In short, our naturalism about
rationality is normative, ecological, and limited. However, it would be a misunderstanding to think that its limitation makes it uninteresting. The cases we shall present
of deriving norms from empirical studies of human reasoning will be striking.
We start with an example concerning investments (Sect. 1) before introducing the
crucial psychological notion of heuristics and explaining how heuristics should be
studied empirically (Sect. 2). Then, we examine the relation between heuristics and
the environments in which we reason. Here we shall also determine the limits of the
present naturalism by sketching central features of the kinds of environments in which
heuristics succeed over standard optimization norms such as the ones embodied in
probability or decision theory (Sect. 3). Furthermore, we explain the basic reasons for
why and when heuristics are to be preferred over optimization strategies that require
more information and calculation. We distinguish two quite different reasons for why
one should rely on heuristics (Sect. 4). Finally, we present and reject an objection
that might easily arise, namely that we are confusing issues of the justification and
application of norms, or that our normative naturalism is built upon ambiguities in the
concept of rationality (Sect. 5).
123
246
the optimal rule for allocating wealth across assets, known as the mean-variance portfolio, because it maximizes the return (mean) and minimizes the risk (variance). Yet
when he made his own retirement investments, instead of using his award-winning
optimization technique, he relied instead on a simple heuristic, the 1/N rule:
Invest equally in each of the N alternatives.
Markowitz was not the only one to use this heuristic; empirical studies indicate
that about 50% of ordinary people intuitively rely on it (Huberman and Jiang 2006).
Skeptics may find this rule naive, however, and ask whether optimization is not the
better alternative. To answer such questions, a study compared the 1/N rule with
the mean-variance portfolio and 13 other optimal asset allocation policies in seven
investment problems, such as allocating ones money to 10 American industry funds
(DeMiguel et al. 2009). The optimization models included sophisticated Bayesian
and non-Bayesian models, which were given 10 years of stock data to estimate their
parameters and on that basis had to predict the next months performance; after this,
the 10-year window was moved a month ahead and the next month had to be predicted,
and so on, until no data remained. 1/N , in contrast, ignores all past information. The
performances of all strategies were evaluated by three standard financial measures:
1/N ranked first on certainty equivalent returns, second on turnover, and fifth on the
Sharpe ratio, respectively. Despite complex estimations and computations, none of the
optimization methods could consistently earn better returns than the simple heuristic.
How can a simple heuristic outperform optimization strategies? Note that in an ideal
world where the mean-variance portfolio could estimate its parameters perfectly, that
is, without error, it would do best. But in an uncertain world, even with 10 years of
data (which is more than the typical investment firm relies on), optimization no longer
necessarily leads to an optimal outcome. In an uncertain world, ignoring information
leads to better decisions. Yet our point is neither that simple heuristics, or satisficing
(Simon 1990), are always better than optimization methods nor that optimization methods are always better, as is typically assumed. No heuristic or optimization strategy is
superior in all worlds. The question instead is in what environment is a given heuristic
more accurate than an optimization strategy, and when is the opposite true? This is the
question of the ecological rationality of a heuristic (Gigerenzer and Selten 2001). The
answer requires analyzing structures of the environment and the match between heuristic and environment. For the choice between 1/N and the mean-variance portfolio,
three relevant environmental features are:
(1) degree of uncertainty,
(2) number N of alternatives, and
(3) size of the learning sample.
It is difficult to predict the future performance of funds, because uncertainty is high.
The size of the learning sample is the estimation window, with 510 years corresponding to the typical calibration of portfolio models in investment practice. 1/N tends to
outperform the mean-variance portfolio if uncertainty is high, the number of alternatives is large, and the learning sample is relatively small. This qualitative insight leads
to a quantitative question: If there are 50 alternatives, how large a learning sample is
needed for the mean-variance portfolio to eventually outperform the simple heuris-
123
247
tic? The answer is: 500 years of stock data, assuming that in the year 2500 the same
fundsand the stock marketstill exist (DeMiguel et al. 2009).
This analysis allows the following statement: The intuitive 1/N heuristic is normatively preferable to the mean-variance optimization method (in the sense that it
leads to better financial outcomes) for N = 50 if there is less than 500 years of stock
data. What is normatively required is no longer defined by consistency principles,
optimization, or armchair thinking but is based on an empirical analysis. It is also not
a statement about what is best, only what is better than alternatives.
2 What heuristics are
As 1/N illustrates, a heuristic is a strategy that ignores some information and minimizes computation, which is exactly what is needed in this particular example for
better (and faster) judgments. The repertoire of heuristics, their building blocks, and
the evolved capacities they exploit can be described as the minds adaptive toolbox
(Gigerenzer 2008).
In Simple heuristics that make us smart (Gigerenzer et al. 1999), a research program was presented that connects the empirical and the normative studies of human
rationality. It has three research goals:
(1)
(2)
(3)
The heuristics in the adaptive toolbox What are the heuristics, their building
blocks, and the evolved capacities in the adaptive toolbox of an individual or a
species?
The connection between heuristics and environment What environmental structures can a heuristic exploit?
Improving reasoning and deciding How can the answers to (1) and (2) be utilized
to design environments and create expert systems that improve judgment and
decision making?
123
248
concluded that when people choose between alternatives, from apartments to contraceptive pills, there is overwhelming evidence for noncompensatory heuristics, whereas
evidence for weighting-and-adding strategies is restricted to tasks with small numbers
of alternatives and attributes (Ford et al. 1989). Recent studies have provided further
evidence for heuristics that make no trade-offs in consumer choice (Kohli and Jedidi
2007; Yee et al. 2007). Studies also showed that when people make inferences, their
reaction times and behavior can be predicted by take-the-best and similar heuristics
typically better than by competing strategies, in particular in situations when relying
on one good reason is adaptive (Bergert and Nosofsky 2007; Brder and Gaissmaier
2007; Nosofsky and Bergert 2007; Rieskamp and Otto 2006). In one area, however,
evidence remains scarce, in part because of the complete dominance of weightingand-adding theories. In choices between gambles, the evidence for one-good-reason
heuristics such as the priority heuristic (Brandsttter et al. 2006) is under debate (e.g.,
Brandsttter et al. 2008; Johnson et al. 2008). Note that not all heuristics are noncompensatory. Tallying, for instance, makes trade-offs. This is a strategy named after
the tally sticks for counting, which can be traced back some 30,000 years in human
history. It uses equal weights (as in 1/N ) and can predict almost as accurately asand
often better thanmultiple linear regression (see Dawes and Corrigan 1974; Dawes
1979; Einhorn and Hogarth 1975; Czerlinski et al. 1999; Schmidt 1971).
The toolbox of heuristics is described as adaptive in order to express the idea that
our reasoning strategies have an evolutionary background. Human beings, therefore,
share at least some heuristics with other animals. The degree to which species share
heuristics will depend on whether they share the evolved capacities on which a heuristic feeds. For instance, humans are genetically disposed to produce language and
provided with cultural input for learning a particular language. It is through language
that legal and social defaults can be communicated, for instance, that by French law
every citizen is a potential organ donor, unless he or she opts out, whereas in the UK
no citizen is a donor, unless he or she opts in. Since most citizens in both countries
rely on the same default heuristic (if there is a default, do nothing, see Table 1), the
result are strikingly different rates of organ donors, 99.9% in France compared to only
16% in the UK (Johnson and Goldstein 2003). A heuristic such as the default heuristic
requires evolved capacities that are unique to humans, and are thus not shared with
other animals.
Whereas language production is not in the adaptive toolbox of animals, or only
in very rudimentary forms, we share other core capacities with many species, such
as object tracking and recognition memory. A shared capacity between two species
makes it likely that these rely on similar heuristics, even if they have to solve different
problems, such as intercepting prey as opposed to fly balls. If the problem is the same
but evolved capacities do not match, different heuristics are required. Consider estimation of area. Ants, for instance, can produce pheromone trails. To measure the area
of a candidate nest cavity, a narrow crack in a rock, an ant has no yardstick but a rule of
thumb based on this capacity: Run around on an irregular path for a fixed period, and
lay down a pheromone trail. Then leave. Return to the cavity, move around on a different irregular path, and estimate its size by the frequency of encountering the old trail.
This heuristic is remarkably precise: Nests half the area of others yielded reencounter
frequencies 1.96 times greater (Mugford et al. 2001). Many of these evolved rules
123
Take-the-best (Gigerenzer
and Goldstein 1996)
Definitiona
Heuristic
Table 1 Ten well-studied heuristics for which there is evidence that they are in humans adaptive toolbox
123
123
a For formal definitions, see references. Each heuristic used to solve problems in social and nonsocial environments. The conditions concerning the ecological rationalITY
and predictions are explained in the references given
Definitiona
Heuristic
Table 1 continued
250
Synthese (2012) 187:243268
251
of thumb in animals, some of which can be found in humans as well, are amazingly
simple and efficient (see the overview by Hutchinson and Gigerenzer 2005). Via their
building blocks based on evolved abilities, they process the patterns of information
available from the environment to produce goal-directed behavior.
123
252
123
253
that emerged during the Enlightenment (Daston 1988; Gigerenzer et al. 1989). Benjamin Franklin described the ideal of weighting and adding all reasons in a letter to
his nephew:
April 8, 1779
If you doubt, set down all the Reasons, pro and
con, in opposite Columns on a Sheet of Paper,
and when you have considered them two or
three Days, perform an Operation similar to
that in some questions of Algebra; observe
what Reasons or Motives in each Column are
equal in weight, one to one, one to two, two
to three, or the like, and when you have struck
out from both Sides all the Equalities, you
will see in which column remains the Balance.
[] This kind of Moral Algebra I have often
practiced in important and dubious Concerns,
and tho it cannot be mathematically exact, I
have found it extremely useful. By the way, if
you do not learn it, I apprehend you will never
be married.
I am ever your affectionate Uncle,
B. Franklin
This vision has inspired the view that complex problems need to be solved by complex algorithms, and that more information is always better. Carnap (1947) proposed
the principle of total evidence, which is the recommendation to use all the available evidence when estimating a probability. Similarly, the statistician Good (1967)
argued that it is irrational to leave observations in the record but not use them. Modern
versions of Franklins moral algebra include expected utility maximization in economics, Bayesian inference theories in the cognitive sciences, Markowitzs mean-variance
optimization model in finance, various bookkeeping principles taught in MBA courses
and recommended by consulting firms, and much of the theories of rationality taught
in philosophy curricula.
Note that Franklin ends with the advice that learning his moral algebra is vital for
marriage. To check, we asked a sample of economists who teach the modern versions
of this theory whether they had chosen their partner using their favorite method. Only
one said that he had. He explained that he listed all the options available and all the
important consequences that he could think of, such as whether she would remain
interesting to talk to after the honeymoon excitement was over, take care of the children, and let him work in peace. He took several days time to estimate the utilities
of each of these consequences, as well as the probabilities that they might actually
occur for each woman. After calculating the expected utility for each candidate, he
proposed to the woman with the highest valuewithout telling her how he had made
his choiceand she accepted. They are now divorced.
The point of this story is emphatically not that Franklins bookkeeping method
produces more divorces than a simple imitate-the-majority heuristic (Table 1): Try
123
254
to get the mate that your peers desire. Rather, our point is that there is a discrepancy between normative theory and actual practice: Despite its being advertised as the
rational method, even devoted proponents often instead rely on heuristics in important decisions. Markowitz is no exception. Reinhard Selten, another Nobel laureate in
economics, who made the chain store problem famous by proving that an aggressive policy against competitors is futile, found his logically correct proof intuitively
unconvincing and said that in the real world, he would rather follow his gut feeling
to be aggressive (Gigerenzer 2007). Many philosophers claim that we should view
norms of rationality as being justified by what Goodman (1965) and Rawls (1971)
have called the reflective equilibrium between principles and intuitions: Whereas
norms can be derived neither from mere descriptions of human behavior (or particular
intuitions about what to do) nor from some metaphysical cloud-cuckoo land, normative
theories can be constructed by striking a balance between a comprehensive amount of
intuitions and principles. To strengthen the equilibrium even more, the experts best
theories should be included in it (Daniels 1979). But as the case of the two Noble
laureates shows, experts often do not follow their own theories, and for a good reason:
Any theory of rationality is ecologically rational in specific environments, not in all
possible worlds.
Health is another example in point. In a study, more than 100 economists were asked
how they decided whether or not to take a prostate-specific antigen (PSA) test (Berg
et al. 2011). Virtually all medical societies recommend that patients carefully weight
pros and cons of PSA testing, since its benefit remains controversial (it is not proven
that it saves lives) but its harms are documented (such as incontinence and impotence
from operations following positive tests). Yet two thirds of the economists said that
they had not weighted any pros and cons but did whatever their doctors or their wives
said they should do. Here, the social heuristic trust your doctor guides important decisions. Note that our point here is descriptive, not evaluative: We question whether the
bookkeeping method, understood as a norm, is justified and contend that it is neither
good nor bad per se. In terms of the normative question, we propose to ask in what
environment a given strategy (such as moral algebra) is better. For instance, in a world
where doctors practice defensive decision making for fear of malpractice lawsuits and
where most doctors do not have the time and education to read the relevant medical
studies but are informed by representatives from the pharmaceutical industry, it pays
to carefully weigh benefits and risks rather than rely on the trust-your-doctor heuristic
(Gigerenzer et al. 2007).
123
255
123
256
123
257
The key disadvantage to the routine use of optimization techniques is that the problem
in question often needs to be replaced by a mathematically convenient small-world
problem (Savage 1972) that can be optimized. For instance, if building an optimization
model for chess, one would have to make it tractable in the first place, for instance, by
shrinking the 8 8 board into a 4 4 board, with a correspondingly reduced set of
figures. Whether the optimal strategy found in the miniature game is also the optimal
one in the real problem, however, remains open. As the case of the mean-variance
portfolio illustrates, an optimization model does not imply optimal behavior in the
real world.
3.2.3 Sample size
In general, the smaller the sample size, the greater the advantage for simple heuristics.
One of the reasons is that complex statistical models have to estimate their parameter
values from past data, and if the sample size is small, then the resulting error due to
variance may exceed the error due to a heuristics bias (Gigerenzer and Brighton
2009). What a small sample size is depends on the degree of uncertainty, as can be seen
in the investment problem where uncertainty is high and a sample size of hundreds of
years of stock data is needed to match and surpass the accuracy of 1/N .
There are also other relevant structures of environments, for instance, redundancy
and variability:
Redundancy Redundancy of cues is often measured by the average correlation between
the cues. Redundant cues can be exploited by sequential heuristics such as take-thebest that rely on the first good reason that allows a decision to be made (Dieckmann
and Rieskamp 2007).
Variability The variability of cue weights can be exploited by heuristics that stop
search quickly. For instance, when variability is high, heuristics that rely on the best
cue alone perform better than when variability is low (Martignon and Hoffrage 2002;
Hogarth and Karelaia 2005, 2006).
Note that our use of the term environment is not identical with the physical or
real environment (Todd 2001). For instance, the degree of uncertainty reflects the
environment (ontic uncertainty) as well as the minds limited understanding (epistemic uncertainty); hence, the degree of uncertainty is located in the system mindenvironment. Similarly, the number of alternatives and the sample size depend on both
what is available in an environment and what an actor or investor actually included
in the consideration set (such as the number N of funds in the analysis). Finally,
redundancy and variability of cues depend both on what is in the physical environment and on humans who select certain cues and not others, which can result
in more or less redundant cues. Physicians, for instance, tend to look at redundant
cues, but they could choose to analyze independent cues and thus in part create their
environment.
As important as physical or biological environments are social environments. For
instance, although an investment can be decided upon individually and made on the
Internet without interacting with other people, the stock market itself is driven by both
nature (e.g., a disastrous hurricane) and people (e.g., the advent of securitization, that
is, the bundling and selling of mortgages). Each of the heuristics in Table 1 can be used
123
258
for social objects (e.g., whom to hire, to trust, to marry) as well as for physical objects
(e.g., what goods to buy). For example, the recognition heuristic exploits environmental structures in which lack of recognition is valuable information and aids inferences
about whom to talk to and whom to trust (dont talk to strangers), but also about in
which stocks to invest. Similarly, satisficing can assist in choosing a mate (Todd and
Miller 1999), but also in buying a pair of trousers, and 1/N can help parents to allocate
their love and time equally to their children, creating a sense of fairness, but also help
investors to diversify.
4 The normative issue: why should people rely on heuristics?
Finally, what are the reasons in favor of using heuristics over optimization strategies
from a normative point of view? To answer this question, we provide critical considerations of why optimization strategies are often inadequate guides for problem solving
and give two reasons in favor of heuristics. This will also reveal that there are (at least)
two different ways of naturalizing the kind of rationality under discussion here. We
begin with experts.
4.1 What strategies do experts rely on?
A first argument against the rationality of simple heuristics is based on widespread
assumptions about the behavior of experts. They are believed to always use optimizing strategies; how could they otherwise be counted as experts? As in the case of
Markowitz and the other examples given in Sect. 2, empirical studies have shown
that experts typically know to distinguish between a heavenly optimization model and
what works in the real world. In general, experts use less information than novices;
they know where to look and tend to rely on limited search and simple heuristics
(Ettenson et al. 1987; Dhami 2003; Dhami and Ayton 2001; Shanteau 1992). This
is illustrated by a study on burglary in which graduate students were given pairs
of residential properties described by eight binary cues, such as flat/house, letter
box empty/stuffed with post, and burglar alarm system yes/no (Garcia-Retamero and
Dhami 2009). The students were asked which property was more likely to be burgled. Two models of cognitive processes were tested, weighting-and-adding of all
information and take-the-best, a simple lexicographic heuristic that bases its decision
on the most important cue only (Table 1). The result was that 78% of the students
were classified as relying on weighting-and-adding and 22% on take-the-best. Normally, psychological experiments stop here. Yet the authors then studied experts,
namely convicted burglars from a male English prison, who reported having successfully committed burglary on an average of 57 occasions. Burglars decisions were
strikingly different: 72% could be predicted by take-the-best and 28% by weightingand-adding. A second expert group, police officers who had investigated residential
burglaries, showed similar reliance on take-the-best. These findings are consistent
with other studies concluding that experienced experts such as managers and physicians tend to rely on simple heuristics (Gigerenzer 2007; Wbben and Wangenheim
2008).
123
259
123
260
say why or simply desires the person that everyone else admired, is the expected basis
of choice for young Western men and women.
4.2.4 Novelty and surprise
An optimization method typically optimizes on past events, and if the future is not
like the past, the methods usefulness is limited in yet another way. We distinguish
three types of uncertainty in prediction. The first is out-of-sample prediction, where
one knows a sample of events in a population, and one has to make predictions about
another sample. This corresponds to the investment problem, where the performance
of funds up to a particular time is known, and one has to predict their performance
in the next month, assuming the market is stable. With this first kind of uncertainty,
optimization methods can already fail to produce better results than a simple heuristic,
depending on the conditions. The second type of uncertainty incurs in out-of-population prediction, where one has information about one population and needs to draw
conclusions about another population that differs in unknown ways. For instance, when
a diagnostic system for heart attacks is validated on a sample of patients in Boston and
applied to patients in Michigan, we face out-of-population prediction. Here, robustness
becomes even more important than in out-of-sample prediction and can be reached by
radically simplifying a diagnostic system, such as replacing a logistic regression with
fast-and-frugal trees (Green and Mehr 1997; Martignon et al. 2003). The third type of
uncertainty is what we call novelty or surprise, where unlike in the other two cases,
all alternatives and consequences are not known. For instance, due to global warming,
more and more animal species living in the northern hemisphere are surprised by new
predators moving north. That is, unforeseen players emerge. To be prepared for novelty,
coarse behavior that appears to be rigid and inflexible has been shown to be superior
to behavior that is fine-tuned and optimized to a past environment that is assumed to
be eternally stable (Bookstaber and Langsam 1985). In other words, to be prepared
for novelty, humans and other animals need to deviate from the optimality conditions
and show behavior that has a chance of surviving when environments change.
To summarize, contrary to the widespread use of optimization in theory (as opposed
to actual expert practice), there are several good reasons not to rely routinely on this
technique for understanding the oughts and is of human behavior (Gigerenzer
2004; Selten 2001).
4.3 Why should people rely on heuristics? Two kinds of arguments (and heuristics)
The classical justification for heuristics is that people save effort, albeit at the cost
of accuracy. There is excellent work showing that, depending on the structure of the
environment, the loss of accuracy can be small (Payne et al. 1993). This interpretation
of the function of heuristics is known as:
Accuracyeffort trade-off Humans and other animals rely on heuristics because
information search and computation cost time and effort; thus, they trade off
some loss in accuracy for faster and more frugal cognition.
123
261
This view starts from the dictum that more is always better, but since information
is not free in the real world, so the argument goes, there is a point at which the costs of
further search exceed the benefits. It underlies optimization-under-constraints theories
in which search in the world (e.g., Stigler 1961) or in memory (e.g., Anderson 1990)
is terminated when the expected costs exceed its benefits. Similarly, the analysis of
the adaptive decision maker (Payne et al. 1993) is built around the assumption that
heuristics achieve a trade-off between accuracy and effort, where effort is a function
of the amount of information and computation consumed. And indeed it is correct that
heuristics often save effort. But with this interpretation, we do not need to revise the
normative assumption that more information and computation are better than less.
The major discovery of recent research on heuristics, however, is that saving effort
does not always lead to a loss in accuracy. Heuristics can be faster and more accurate
than strategies that use more information and more computation. Our analysis of the
function of heuristics goes beyond accuracyeffort trade-offs:
Less can be more Humans and other animals rely on heuristics in situations where
these are ecologically rational, including situations where less information and
computation lead to more accurate judgments.
Here is the essential difference. The interpretation that relying on a heuristic is functional because of an accuracyeffort trade-off is based on gold standards of rationality,
including weighting-and-adding models in the spirit of Franklins moral algebra and
Bayes rule. By its very definition, no heuristic can be more accurate. In contrast, the
study of ecological rationality treats the normative question as an empirical one and
allows and investigates less-is-more effects, as in the investment example.
Less-is-more effects are surprising indeed. One explanation for them is that a strategy can make two kinds of errors: bias and variance. Specialized tools such as heuristics tend to have larger bias than more general ones, and general-purpose tools with
many free parameters tend to generate more variance. A good cognitive system needs
strategies that strike a balance between being too specialized and being too general.
From this it follows that a mind with a general-purpose algorithm would not only be
slow, but also perform inferior to a more specialized system if the error due to variance
was larger than the error due to bias (for details, see Brighton and Gigerenzer 2008;
Gigerenzer and Brighton 2009).
Less-is-more effects have been studied and demonstrated for some time, such as
in the seminal work by Dawes and colleagues on tallying (Dawes and Corrigan 1974;
Dawes 1979) and on take-the-best (Gigerenzer and Brighton 2009; Gigerenzer and
Goldstein 1996). Yet relevant studies have been ignored for decades (Makridakis and
Hibon 2000). When issues of rationality are at stake, emotions can arise that are resistant to evidence. If you open a book on cognition or decision making, the following
three misconceptions are still likely to be asserted:
Misapprehension 1 Heuristics are second-best; optimization is always better.
Misapprehension 2 We use heuristics only because of our cognitive limitations.
Misapprehension 3 More information, more computation, and more time are always
better.
123
262
Each statement stems from the assumed generality of the accuracyeffort trade-off.
Yet each is incorrect as a general belief, and more and more studies document less-ismore effects. For instance, in an age in which companies maintain databases of their
customers, complete with historical purchase data, a key question becomes predicting which customers are likely to purchase again in a given time frame, and which
are inactive. Wbben and Wangenheim (2008) found that managers in airline and
apparel industries rely on a simple hiatus heuristic: If a customer has not purchased
within 9 months in the past (the hiatus), the customer is classified as inactive, and
otherwise active. They compared the hiatus heuristic with the Pareto/NBD model,
which assumes that purchases follow a Poisson process with a purchase rate parameter , that customer lifetimes follow an exponential distribution with a dropout rate
parameter , and that, across customers, purchase and dropout rates are distributed
according to a gamma distribution. For both industries, the simple heuristic correctly
classified more customers than the Pareto/NBD model. Similarly, in library search, a
one-reason decision heuristic produced better order of titles than a Bayesian model
and PsychInfo (Lee et al. 2002). For many inference problems in the real world, there
is an inversely U-shaped relation between amount of information, computation, and
time on the one hand and accuracy on the other.
There are thus two quite different reasons for why people doand ought torely
on heuristics: For one, there are many tasks where a trade-off is made between accuracy and effort. At the same time, there are also many tasks where a less-is-more
effect is visible. In the first of these cases, heuristics are to be recommended in order
to save time or energy, thereby leaving open whether there are other norms that, given
only more time or energy, would lead to more accurate or better results. In the second
case, however, it would be irrational to apply norms that require more information or
computation. Psychology can thus be used to derive norms in two different ways: first,
in order to save time or other resources, and second, in order to refute the widespread
universal claim that more information and computation always lead to better results.
Experiments show that they do not.
123
263
events cannot exceed the probability of each conjunct, is a justified truth of the probability calculus. Yet, it may be inappropriate to apply it as a norm in certain reasoning
tasks (as discussions of the notorious Linda problem have clarified; see Hertwig
and Gigerenzer 1999, with reference to Grice et al.s [1975] conversational maxim
of relevance). More generally, formal rules cannot simply be equated with norms of
rationality (Goldman 1986, 2008).
Now, the objection continues, if a particular law or rule should not be applied in
certain situations, this hardly implies that the law or rule in question is not perfectly
justified or that ideal reasoners should not use it. A variety of such thinking occurs
in ethical discussion about utilitarianism: It may be difficult to identify the rule or
action that is optimal in the sense that it maximizes happiness or expected utility, yet
that does not undermine the highest principle of utilitarianism. We can distinguish
between practical and ideal norms (Birnbacher 1988). It is obvious how such an
argument might be transferred to the debate about the normative evaluation of heuristics, at least those that are based on accuracyeffort trade-offs: They fall short of some
optimization strategy, which itself may be justified by a comprehensive normative theory of rationality. There are various ways in which the objection can be formulated,
related to different meanings and uses of the term rational, some of which we hinted
at in the introduction. For instance, one might say that such heuristics are normatively
required only in practice, not in principle. Alternatively, one might say that we need to
distinguish between the rationality of rules and the rationality of agents or their inferences and decisions. According to the objection starting from the distinction between
justification and application, we are confusing the question of whether a norm is justified with the question of the case-sensitive application of the norm. Alternatively, our
position trades on ambiguities in the concept of rationality. Does that not undermine
our naturalization of rationality?
We see two main problems with this objection. First, the argument delivers merely a
promise. It has to be shown in each and every case where a trade-off between accuracy
and effort is made that there is indeed an optimization strategy that is theoretically
better justified. And even if such proof were given, there would often be good, namely
practical reasons for not using the optimization strategy in the real world. Second, and
more importantly, the argument does not work for less-is-more effects. Here, a simple
heuristic can be equally or more accurate than a strategy that relies on more information and computation, including optimization methods. Hence, no formal system, from
versions of logic to those of Bayesianism, can be justified as the rational solution for
all problems. Savage (1972), who is credited as the father of modern Bayesian decision
theory, emphasized that this theory can only be justified as a theory of rationality in
a small world where all the relevant pieces of information are known and where no
surprises occur. Yet contemporary Bayesianism often appears to assume that Bayesian
decision theory is a general-purpose tool for rationality in both small and large worlds
(see Binmore 2009). More generally, the application of every strategy, optimization
or heuristic, can be justified only relative to the structure of the environment. In our
view, optimization and heuristic strategies should therefore be treated equally. Since
less-is-more effects exist, the usual distinction between justification and application,
which may have be legitimate elsewhere, breaks down.
123
264
There is a historical irony here. The distinction between justification and application stands in a venerable tradition, it being closely related to Immanuel Kants
distinction between the faculties of reason and judgment. Kant has often been viewed
as a nave defender of abstract reason. He was, however, fully aware that the proper
application of norms is an important task that cannot be achieved by a priori reasoning.
To know when and where and how to correctly apply a certain rule is, as he said, a
matter of a well-trained and experienced judgment (Urteilskraft). Rather than acquiring it through a priori reasoning, one needs to learn to describe a particular situation correctly and to find the one norm among many to be applied to the situation.
Kant even forcefully argued that there cannot be an a priori, strictly general rule
guiding the application of rules to problem situations: Any such rule would in turn
require another rule guiding its application, which leads to an infinite regress (Kant
1781/1787, pp. A132/B171A134/B174). This essay is silent about Kants views on
the existence and function of a priori principles either in his theoretical or practical
philosophy, because the ones he identified and defended have nothing to do with the
norms of optimizing conceptions of rationality, especially in the decision sciences.
Nonetheless, Kant was profoundly right in claiming that factual knowledge about
the (social and the natural) world must be coupled to norms in order to produce
good reasoning. This pertains even to optimization strategies: Their applicability also
depends on whether they fit with certain environments. Moreover, when a less-is-more
heuristic can help to solve problems for which these optimization strategies were originally designed, the very point of their justification within a certain system becomes
unclear.
To sum up: We doubt that the justification-application distinction could be used to
undermine our normative naturalism. Like Kant, we emphasize that there is no single
general-purpose rule of good reasoning and decision making but that there is a choice
between several that requires thinking about the structure of the world.
6 Conclusion: heuristics as descriptive and normative
The classical view of heuristics as lazy shortcuts, falling short of some idealized model
of general rationality, grants the study of heuristics only a descriptive role. It draws a
strict line between how behavior is and how it should be, with psychology answering
the first question but silenced on the second, which is said to be in the territory of normative epistemology or decision theory conceived of as nonempirical enquiries. The
empirically grounded study of ecological rationality, in contrast, provides a descriptive and prescriptive role for heuristics. In the right environment, a heuristic can be
better than an optimization model or other complex strategies. This is not to say that
optimization strategies have no role at all, but their role is restricted to well-defined
problems with (almost) error-free parameter estimates. In this regard, our naturalism
and its concomitant notion of ecological rationality extend themselves even to the correct application of optimization strategies. Mind and environment hang together, and
a theory of rationality should respect and explicitly deal with both blades of Simons
scissors.
123
265
Acknowledgements The authors are grateful to two anonymous referees for various comments and
suggestions. Thomas Sturms work on this article was supported by the Spanish Ministry for Science and
Innovation, Reference number FFI 2008-01559/FISO.
References
Alexy, R. (1993). Justification and application of norms. Ratio Juris, 6, 157170.
Anderson, J. R. (1990). The adaptive character of thought. Hillsdale, NJ: Erlbaum.
Axelrod, R. (1984). The evolution of cooperation. New York: Basic Books.
Berg, N., Biele, G., & Gigerenzer, G. (2011). Logical consistency and accuracy of beliefs: Survey
evidence on health decision-making among economists. Unpublished manuscript.
Bergert, F. B., & Nosofsky, R. M. (2007). A responsetime approach to comparing generalized rational
and take-the-best models of decision making. Journal of Experimental Psychology: Learning,
Memory, and Cognition, 33, 107129.
Binmore, K. (2009). Rational decisions. Princeton, NJ: Princeton University Press.
Birnbacher, D. (1988). Verantwortung fr zuknftige Generationen. Stuttgart: Reclam.
Bishop, M. (2008). Reflections on a normative psychology. In A. Beckermann & S. Walter (Eds.),
Philosophy: Foundations and applications (pp. 25262). Paderborn: Mentis.
Bishop, M., & Trout, J. D. (2005). Epistemology and the psychology of human judgment. New
York: Oxford University Press.
Bookstaber, R., & Langsam, J. (1985). On the optimality of coarse behavior rules. Journal of Theoretical
Biology, 116, 161193.
Boyd, R., & Richerson, P. J. (2005). The origin and evolution of cultures. New York: Oxford University
Press.
Brandsttter, E., Gigerenzer, G., & Hertwig, R. (2006). The priority heuristic: A process model of risky
choice. Psychological Review, 113, 409432.
Brandsttter, E., Gigerenzer, G., & Hertwig, R. (2008). Risky choice with heuristics: Reply to Birnbaum
(2008), Johnson, Schulte-Mecklenbeck, and Willemsen (2008) and Rieger and Wang (2008).
Psychological Review, 115, 281290.
Brighton, H. (2006). Robust inference with simple cognitive models. In C. Lebiere & R. Wray (Eds.),
Between a rock and a hard place: Cognitive science principles meet AI-hard problems. Papers
from the AAAI Spring symposium (AAAI Tech. Rep. No. SS-06-03, pp. 1722). Menlo Park, CA:
AAAI Press.
Brighton, H., & Gigerenzer, G. (2008). Bayesian brains and cognitive mechanisms: Harmony or dissonance?. In N. Chater & M. Oaksford (Eds.), The probabilistic mind: Prospects for Bayesian
cognitive science (pp. 189208). New York: Oxford University Press.
Brder, A. (2003). Decision making with the adaptive toolbox: Influence of environmental structure,
intelligence, and working memory load. Journal of Experimental Psychology, 29, 611625.
Brder, A., & Gaissmaier, W. (2007). Sequential processing of cues in memory-based multi-attribute
decisions. Psychonomic Bulletin and Review, 14, 895900.
Bruss, F. T. (2000). Der Ungewissheit ein Schnippchen schlagen. Spektrum der Wissenschaft, 6, 106107.
Carnap, R. (1947). On the application of inductive logic. Philosophy and Phenonmenlogical
Research, 8, 133148.
Chiappe, D. L., & Vervaeke, J. (1997). Fodor, Cherniak, and the naturalization of rationality. Theory
& Psychology, 7, 799821.
Cooper, G. F. (1990). Computational complexity of probabilistic inference using Bayesian belief
networks. Artificial Intelligence, 42, 393405.
Czerlinski, J., Gigerenzer, G., & Goldstein, D. G. (1999). How good are simple heuristics? In G.
Gigerenzer, P. M. Todd, & the ABC Research Group (Eds.), Simple heuristics that make us smart
(pp. 97118). New York: Oxford University Press.
Dagum, P., & Luby, M. (1993). Approximating probabilistic inference in Bayesian belief networks is
NP-hard. Artificial Intelligence, 60, 141153.
Daniels, N. (1979). Wide reflective equilibrium and theory acceptance in ethics. The Journal of
Philosophy, 76, 256282.
Daston, L. J. (1988). Classical probability in the enlightenment. Princeton, NJ: Princeton University
Press.
123
266
Dawes, R. M. (1979). The robust beauty of improper linear models in decision making. American
Psychologist, 34, 571582.
Dawes, R. M., & Corrigan, B. (1974). Linear models in decision making. Psychological Bulletin, 81, 95
106.
DeMiguel, V., Garlappi, L., & Uppal, R. (2009). Optimal versus naive diversification: How inefficient
is the 1/N portfolio strategy? Review of Financial Studies, 22, 19151953.
Dhami, M. K. (2003). Psychological models of professional decision-making. Psychological Science,
14, 175180.
Dhami, M. K., & Ayton, P. (2001). Bailing and jailing the fast and frugal way. Journal of Behavioral
Decision Making, 14, 141168.
Dieckmann, A., & Rieskamp, J. (2007). The influence of information redundancy on probabilistic
inferences. Memory & Cognition, 35, 18011813.
Dudey, T., & Todd, P. M. (2002). Making good decisions with minimal information: Simultaneous and
sequential choice. Journal of Bioeconomics, 3, 195215.
Einhorn, H. J., & Hogarth, R. M. (1975). Unit weighting schemes for decision making. Organizational
Behavior and Human Performance, 13, 171192.
Ettenson, R., Shanteau, J., & Krogstad, J. (1987). Expert judgment: Is more information better?. Psychological Report, 60, 227238.
Ford, J. K., Schmitt, N., Schechtman, S. L., Hults, B. H., & Doherty, M. L. (1989). Process tracing
methods: Contributions, problems, and neglected research questions. Organizational Behavior and
Human Decision Processes, 43, 75117.
Garcia-Retamero, R., & Dhami, M. K. (2009). Take-the-best in expert-novice decision strategies for
residential burglary. Psychonomic Bulletin & Review, 16, 163169.
Gigerenzer, G. (2004). Striking a blow for sanity in theories of rationality. In M. Augier & J. G.
March (Eds.), Models of a man: Essays in memory of Herbert A.Simon (pp. 389409). Cambridge,
MA: MIT Press.
Gigerenzer, G. (2007). Gut feelings: The intelligence of the unconscious. New York: Viking Press.
Gigerenzer, G. (2008). Rationality for mortals. New York: Oxford University Press.
Gigerenzer, G., & Brighton, H. (2009). Homo heuristicus: Why biased minds make better inferences.
Topics in Cognitive Science, 1, 137.
Gigerenzer, G., Gaissmaier, W., Kurz-Milcke, E., Schwartz, L. M., & Woloshin, S. W. (2007). Helping doctors and patients make sense of health statistics. Psychological Science in the Public
Interest, 8, 5396.
Gigerenzer, G., & Goldstein, D. G. (1996). Reasoning the fast and frugal way: Models of bounded
rationality. Psychological Review, 103, 650669.
Gigerenzer, G. & Selten, R. (Eds.). (2001). Bounded rationality: The adaptive toolbox. Cambridge,
MA: MIT Press.
Gigerenzer, G., Swijtink, Z., Porter, T., Daston, L., Beatty, J., & Krger, L. (1989). The empire of
chance. How probability changed science and everyday life. Cambridge: Cambridge University
Press.
Gigerenzer, G., Todd, P. M., & the ABC Research Group. (1999). Simple heuristics that make us smart.
New York: Oxford University Press.
Gilbert, J. P., & Mosteller, F. (1966). Recognizing the maximum of a sequence. American Statistical
Association Journal, 61, 3573.
Goldman, A. (1986). Epistemology and cognition. Cambridge, MA: Harvard University Press.
Goldman, A. (2008). Human rationality: Epistemological and psychological perspectives. In A. Beckermann & S. Walter (Eds.), Philosophy: Foundations and applications (pp. 230247). Paderborn: Mentis.
Goldstein, D. G., & Gigerenzer, G. (2002). Models of ecological rationality: The recognition heuristic. Psychological Review, 109, 7590.
Good, I. J. (1967). On the principle of total evidence. The British Journal for the Philosophy of
Science, 17, 319321.
Goodman, N. (1965). Fact, fiction, and forecast. Indianapolis, IN: Bobbs-Merrill.
Green, L. A., & Mehr, D. R. (1997). What alters physicians decisions to admit to the coronary care
unit?. The Journal of Family Practice, 45, 219226.
Grice, H. P. (1975). Logic and conversation. In P. Cole & J. L. Morgan (Eds.), Syntax and semantics,
vol. 3: Speech acts (pp. 4158). New York: Academic Press.
123
267
123
268
Rieskamp, J., & Otto, P. E. (2006). SSL: A theory of how people learn to select strategies. Journal of
Experimental Psychology: General, 135, 207236.
Savage, L. J. (1972). The foundations of statistics. New York: Dover.
Schmidt, F. L. (1971). The relative efficiency of regression and simple unit weighting predictor weights
in applied differential psychology. Educational and Psychological Measurement, 31, 699714.
Schooler, L. J., & Hertwig, R. (2005). How forgetting aids heuristic inference. Psychological Review,
112, 610628.
Selten, R. (2001). What is bounded rationality?. In G. Gigerenzer & R. Selten (Eds.), Bounded rationality:
The adaptive toolbox (pp. 1336). Cambridge, MA: MIT Press.
Serwe, S., & Frings, C. (2006). Who will win Wimbledon? The recognition heuristic in predicting
sports events. Journal of Behavioral Decision Making, 19, 321332.
Shanteau, J. (1992). How much information does an expert use? Is it relevant?. Acta Psychologica, 81, 7586.
Simon, H. A. (1955). A behavioral model of rational choice. Quarterly Journal of Economics, 69, 99118.
Simon, H. A. (1990). Invariants of human behavior. Annual Review of Psychology, 41, 119.
Smith, V. L. (2003). Constructivist and ecological rationality in economics. American Economic
Review, 93, 465508.
Stein, E. (1996). Without good reason: The rationality debate in philosophy and cognitive science.
Oxford: Clarendon.
Stigler, G. J. (1961). The economics of information. Journal of Political Economy, 69, 213225.
Todd, P. M. (2001). Fast and frugal heuristics for environmentally bounded minds. In G. Gigerenzer
& R. Selten (Eds.), Bounded rationality: The adaptive toolbox (pp. 5170). Cambridge, MA: MIT
Press.
Todd, P. M., & Miller, G. F. (1999). From pride and prejudice to persuasion: Satisficing in mate search.
In G. Gigerenzer, P. M. Todd, & the ABC Research Group (Eds.), Simple heuristics that make
us smart (pp. 287308). New York: Oxford University Press.
Tversky, A. (1972). Elimination by aspects: A theory of choice. Psychological Review, 79, 281299.
Wbben, M., & Wangenheim, F. v. (2008). Instant customer base analysis: Managerial heuristics often
get it right. Journal of Marketing, 72, 8293.
Yee, M., Hauser, J., Orlin, J., & Dahan, E. (2007). Greedoid-based non-compensatory two-stage
consideration-then-choice inference. Marketing Science, 26, 532549.
123
Copyright of Synthese is the property of Springer Science & Business Media B.V. and its content may not be
copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written
permission. However, users may print, download, or email articles for individual use.